Abstract
Abstract In recent years, there are increasing demands for audio-visual communication using low bit rate channels (with bit rates ranging from 10 to 60 kilobits per second.) The process of developing new technology for video compression in the context of the emerging ISO-MPEG-4 standard has resulted in a large amount of work being produced in the framework of motion-compensated prediction-based video coding. All existing video compression standards are hybrid systems in that the compression is archived in two stages: first, motion compensation followed by encoding the residual frame due to the prediction error of motion compensation. In the current video compression standards, such as widely known ISO MPEG-1 and MPEG-2 standards as well as the ITU-T video coding standards H.261 and H.263, block-based discrete cosine transform (DCT) has been used to encode these prediction errors. DCT based video compression scheme is efficient but it introduces undesirable blocking artifacts at low bit rates. Hence, In this thesis, we propose a codebook adaptation algorithm, such that Kiefer-Wolfowitz method, for low bit rate, real-time video compression based on matching pursuit (MP). Although adaptive codebook design has been studied in the past, its implementation at low bit rate coding suitable for MPEG-4 standard remains significantly challenging. In our adaptation algorithm, we use a subset of 2-D separable Gabor function as our initial dictionary. Our initial dictionary is selected the same as Neff and Zakhor's. Also, in order to speed up the convergence rate of our adaptation algorithm, the basis functions in our dictionary is formed by tensor product of x-component and y-component. Each basis consists of two code vectors, one from the x-direction and the other from the y-direction. We start with an initial dictionary and on-line adapt this dictionary suitable for current sequences. Experiments of our adaptive matching pursuit video codec on several MPEG-4 Class A, Class B, and Class C sequences.